# Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck

Source: [Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck](<https://www.sota2.com/research/paper/69b24878d030084c22e2f85c>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-13T05:00:04.336Z

Authors: Hongbin Zhang, Kehai Chen, Xuefen Bai, Youcheng Pan, Yang Xiang, Jinpeng Wang, Min Zhang

Published: 2026-03-11T02:55:29.000Z

## Abstract

Large language models (LLMs) have become a standard for multilingual evaluation, yet they exhibit a severe systematic translationese bias. In this paper, translationese bias is characterized as LLMs systematically favoring machine-translated text over human-authored references, particularly in low-resource languages. We attribute this bias to spurious correlations with (i) latent manifold alignment with English and (ii) cross-lingual predictability. To mitigate this bias, we propose DIBJudge, a robust fine-tuning framework that learns a minimally sufficient, judgment-critical representation via variational information compression, while explicitly isolating spurious factors into the dedicated bias branch. Furthermore, we incorporate a cross-covariance penalty that explicitly suppresses statistical dependence between robust and bias representations, thereby encouraging effective disentanglement. Extensive evaluations on multilingual reward modeling benchmarks and a dedicated translationese bias evaluation suite demonstrate that the proposed DIBJudge consistently outperforms strong baselines and substantially mitigates translationese bias.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Multilingual Reward Modeling on m-RewardBench Avg. 23 langs](<https://www.sota2.com/research/sota/multilingual-reward-modeling-on-m-rewardbench-avg-23-langs.md>): Multilingual Reward Modeling; m-RewardBench Avg. 23 langs; Accuracy: 91.37.

- [Multilingual Reward Modeling on MM-Eval Avg. 18 lang](<https://www.sota2.com/research/sota/multilingual-reward-modeling-on-mm-eval-avg-18-lang.md>): Multilingual Reward Modeling; MM-Eval Avg. 18 lang; Accuracy: 87.53.

- [Reward Modeling on RewardBench English](<https://www.sota2.com/research/sota/reward-modeling-on-rewardbench-english.md>): Reward Modeling; RewardBench English; Accuracy: 91.01.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.10351>)

[PDF](<https://arxiv.org/pdf/2603.10351.pdf>)
